PE-Net: A Plane Extraction Network Generating Plane Constraints for Pose Estimation

Plane, as a common indoor geometric feature, is widely used in robot positioning or navigation. At present, traditional algorithms used to extract planes are limited by sensors and other aspects, while methods such as deep learning methods to extract planes from RGB images are limited to empirical information, and there are some other problems such as uncertain scales. On the other hand, the visual positioning systems based on feature points or the direct method also have the problem of low accuracy when there are few textures. This paper designs a deep convolutional neural network that takes RGB-D (RGB-depth) images as input and performs instance segmentation on planes, and a visual positioning algorithm based on Elastic Fusion and plane constraints to improve camera positioning. The experimental results show that the algorithm that fuses RGB-D information has better effect on plane classification and instance segmentation than related algorithms that simply use color or depth information. The visual positioning algorithm using the extracted plane can improve the positioning accuracy of the camera by the 3D reconstruction algorithm.